The CLASS Software Architecture: Edge Analytics and Use Cases

Page created by Annette Byrd
 
CONTINUE READING
The CLASS Software Architecture: Edge Analytics and Use Cases
COORDINATING EDGE AND CLOUD
                                             FOR BIG DATA ANALYTICS

The CLASS Software Architecture:
Edge Analytics and Use Cases

Roberto Cavicchioli, UNIMORE
HiPEAC Virtual venue, April 22, 2021
    The CLASS project has received funding from the European Union's Horizon 2020 research and innovation programme under the grant agreement No 780622
The CLASS Software Architecture: Edge Analytics and Use Cases
Outline
 The MASA infrastructure
 Use Cases
 Analytics

HiPEAC 2021, Roberto Cavicchioli   2
The CLASS Software Architecture: Edge Analytics and Use Cases
Infrastructure
The CLASS Software Architecture: Edge Analytics and Use Cases
Use Case Infrastructure – Modena Automotive
Smart Area

 HiPEAC 2021, Roberto Cavicchioli
The CLASS Software Architecture: Edge Analytics and Use Cases
On-board vehicle Components
    Blindspot                                                              Surroundings perception
   Radar 120°                                                              - 6 Camera
                                                                           - 3 Radar
                                           120°                            - 1 LiDAR
                           LiDAR 360°
GPS/GNSS                                          60°                      Vehicle positioning
                                                                           - GPS/GNSS receiver

                                                         Front Radar 40°   V2I communication
                    120°                          120°
                                                                           - 4G-LTE antenna

                                                   60°
                       Computation         120°
                        processor
    Blindspot
   Radar 120°

        HiPEAC 2021, Roberto Cavicchioli
The CLASS Software Architecture: Edge Analytics and Use Cases
On-board vehicle Components
Maserati Quattroporte MY19
      Units/car                     Component                     Model

            1         Computation processor     Nvidia board – DRIVE AGX Pegasus

            1         LiDAR (360°)              OUSTER OS2-128

            2         Camera (60°)              Sekonix SF3325-10X

            4         Camera (120°)             Sekonix SF3324-10X

            1         GPS/GNSS                  Xsens MTi-G-710-GNSSINS-2A8G4-DK

 HiPEAC 2021, Roberto Cavicchioli
The CLASS Software Architecture: Edge Analytics and Use Cases
Use Cases
The CLASS Software Architecture: Edge Analytics and Use Cases
Obstacle Detection scenario
Scenario      Location in MASA                      Actors                 Functionality
                                         City camera, Smart Car (SC),
   1        Via Manfredo Fanti -        Connected Car (CC), and Non-       Virtual mirror
                Str. Attiraglio              Connected Car (!CC)
                                       Two City cameras, Smart Car (SC),
   2            Str. Attiraglio                                            Virtual mirror
                                            Truck, and Pedestrian

             Via Manfredo Fanti       Two Smart Cars (SC), Connected Car
   3                                                                       Virtual mirror
             (at the parking exit)            (CC), and Cyclist

           Via Maria Montessori - Two City cameras, Smart Car (SC),        Two sources of
   4       Via Pico della Mirandola          and Cyclist                     attention
             (at the roundabout)
                                       City camera, Connected Car (CC),
           Via Pico della Mirandola                                        Two sources of
   5                                    Non-Connected Car (!CC), and
             (at the parking exit)                                           attention
                                                 Pedestrian

  HiPEAC 2021, Roberto Cavicchioli
The CLASS Software Architecture: Edge Analytics and Use Cases
Obstacle Detection: scenario 2
Virtual Mirror

                                   The actors involved are: two city cameras at
                                   Attiraglio street that detect a pedestrian, a
                                   Maserati car (SC), and a truck that hides the
                                   view of the SC driver. The driver will receive
                                   an alert due to a hazard detection, since the
                                   pedestrian is crossing the street.

HiPEAC 2021, Roberto Cavicchioli
The CLASS Software Architecture: Edge Analytics and Use Cases
Obstacle Detection: scenario 5
2 sources of attention

                                   The actors involved are: a connected car (CC)
                                   exiting the parking, and incorporating to the
                                   roundabout; a city camera detecting a non-
                                   connected car (!CC) and a pedestrian. The CC
                                   will receive an alert various hazard detection.

HiPEAC 2021, Roberto Cavicchioli
Analytics
Position of objects

   Detection performed on camera → we obtain
    bounding boxes on the camera image
   GPS conversion → all the lower centers (pixel x,
    pixel y) of the bounding box are converted in
    (GPS latitude, GPS longitude)
   Meters → the GPS points (GPS latitude, GPS
    longitude) are finally converted in meters (x [m],
    y[m]) w.r.t an origin on order to feed the tracker
    with meters and not GPS coordinates and
    perform tracking from a map point of view and
    not from the camera itself.

HiPEAC 2021, Roberto Cavicchioli                         12
DNN Performance
                    Yolo v3 COCO            Yolo v3 Berkeley Yolo v3 Berkeley Yolo v3 Berkeley Yolo v2 Berkeley Yolo v4 Berkeley
                    (416x416)               (416x416)        (544x320)        (488x256)        (544x320)        (544x320)

FULL                6 fps                   6 fps              5.5 fps           10 fps           11.5 fps           5.6 fps

HALF                7.5 fps                 7.5 fps            7.5 fps           12 fps           17 fps             7.2 fps

Map                                  0,94              44,63             47,07            45,16              33,77              51,8

map person                          70,24              45,81             51,54            47,61              32,04             59,11

map car                              4,78               66,7             70,78             66,4              51,78              76,5

map truck                            0,02              57,34             58,14            57,01              51,23              64,0

map bus                              0,01              55,09             56,65            55,78              51,63              64,2

map bike                             0,05              39,05             41,41            40,53              33,49             40,63

map motor                            0,03               34,3             36,57            39,66               31,2             41,01

 HiPEAC 2021, Roberto Cavicchioli                                                                                                13
Tracking Overview

   Filter every object using the lower center of the bounding box
   Extended Kalman Filter
   Nearest neighbor technique to associate the bounding boxes of the same
    object in successive frames.
   One tracker for each object → multiple trackers

             Extended Kalman       Nearest                 Multiple
                  Filter           Neighbor                trackers
HiPEAC 2021, Roberto Cavicchioli                                             14
Multiple trackers: view from camera

HiPEAC 2021, Roberto Cavicchioli      15
Multiple trackers: view from the top

HiPEAC 2021, Roberto Cavicchioli       16
Thank you!
    Questions?

       www.class-project.eu
        Twitter: @EU_CLASS
LinkedIn: http://bit.ly/CLASS-project
Backup
Dynamic Work Distribution
DAG nodes

 Alternative nodes: to model alternative implementations of
  the same functionality on different computing engines to be
  selected off-line

 Conditional nodes: to model if-then-else branches to be
  selected at run-time

HiPEAC 2021, Roberto Cavicchioli                                22
Scheduling analysis

 Each engine has its own scheduler and a separate ready-queue.
      - Sub-tasks are allocated (partitioned) onto the available engines so that the system is
        schedulable.
 Meeting timing constraints of a concrete task depends on the allocation of the
  sub-tasks onto the different execution engines.
      - To reduce the complexity of dealing with precedence constraints directly,
        we impose intermediate offsets and deadlines on each sub-task.

HiPEAC 2021, Roberto Cavicchioli                                                                 23
Task Allocation

 ILP formulation of this problem fails to produce feasible solutions in an
  acceptable short time, therefore we propose a set of greedy heuristics to quickly
  explore the space of solutions.
      - Single-engine allocation
      - Best-fit/Worst-fit allocation
      - Serial/Parallel allocation

HiPEAC 2021, Roberto Cavicchioli                                                  24
Obstacle Detection: scenario 1

                                   The actors involved are: a city camera at
                                   Attiraglio street that detects a vehicle (!CC),
                                   and a Maserati car (SC) that detects a
                                   connected car (CC) at Via Manfredo Fanti,
                                   both reaching the intersection. The connected
                                   car (CC) will receive an alert due to a hazard
                                   detection, since the non-connected car will exit
                                   Attiraglio street.

HiPEAC 2021, Roberto Cavicchioli
Obstacle Detection: scenario 3

                                   The actors involved are: a Maserati car (SC 1)
                                   exiting the parking; and a second Maserati car
                                   (SC 2) that detects a connected car (CC) at
                                   Via Manfredo Fanti, and a cyclist, the three of
                                   them reaching the intersection with the parking
                                   exit. Both, the SC 1 and the CC, will receive
                                   an alert due to various hazard detection.

HiPEAC 2021, Roberto Cavicchioli
Obstacle Detection: scenario 4

                                   The actors involved are: a connected car (CC)
                                   reaching the roundabout; and two city
                                   cameras that detect the CC and a Cyclist at
                                   the roundabout. The CC will receive an alert
                                   due to a hazard detection, since the cyclist
                                   may be crossing the street.

HiPEAC 2021, Roberto Cavicchioli
On-board vehicle Components (2/4)
Maserati Quattroporte MY18
       Units/car                    Component                     Model

            1         Computation processor     Nvidia board - DRIVE PX2 Autochauffeur

            1         LiDAR (360°)              VLP-16 channel PUCK

            2         Camera (60°)              Sekonix SF3325-10X

            4         Camera (120°)             Sekonix SF3324-10X

            1         GPS/GNSS                  Xsens MTi-G-710-GNSSINS-2A8G4-DK

 HiPEAC 2021, Roberto Cavicchioli
On-board vehicle Components (3/4)
Maserati Levante
        Units/car                  Component                  Model

             1         Computation processor   Nvidia board - DRIVE AGX Xavier

             1         LiDAR (360°)            OUSTER OS1-64

             2         Camera (60°)            Leopard Imaging LI-AR0231-GMSL

             4         Camera (120°)           Leopard Imaging LI-AR0231-GMSL

             1         GPS/GNSS                Xsens MTi-G-710-GNSSINS-2A8G4-DK

HiPEAC 2021, Roberto Cavicchioli
You can also read